1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Implementation of cl_lock for OSC layer.
40 * Author: Nikita Danilov <nikita.danilov@sun.com>
43 #define DEBUG_SUBSYSTEM S_OSC
46 # include <libcfs/libcfs.h>
48 # include <liblustre.h>
50 /* fid_build_reg_res_name() */
51 #include <lustre_fid.h>
53 #include "osc_cl_internal.h"
59 #define _PAGEREF_MAGIC (-10000000)
61 /*****************************************************************************
67 static const struct cl_lock_operations osc_lock_ops;
68 static const struct cl_lock_operations osc_lock_lockless_ops;
69 static void osc_lock_to_lockless(const struct lu_env *env,
70 struct osc_lock *ols, int force);
71 static int osc_lock_has_pages(struct osc_lock *olck);
73 int osc_lock_is_lockless(const struct osc_lock *olck)
75 return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
79 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
80 * pointer cannot be dereferenced, as lock is not protected from concurrent
81 * reclaim. This function is a helper for osc_lock_invariant().
83 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
85 struct ldlm_lock *lock;
87 lock = ldlm_handle2lock(handle);
94 * Invariant that has to be true all of the time.
96 static int osc_lock_invariant(struct osc_lock *ols)
98 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
99 struct ldlm_lock *olock = ols->ols_lock;
100 int handle_used = lustre_handle_is_used(&ols->ols_handle);
103 ergo(osc_lock_is_lockless(ols),
104 ols->ols_locklessable && ols->ols_lock == NULL) ||
105 (ergo(olock != NULL, handle_used) &&
107 olock->l_handle.h_cookie == ols->ols_handle.cookie) &&
109 * Check that ->ols_handle and ->ols_lock are consistent, but
110 * take into account that they are set at the different time.
113 ergo(lock != NULL && olock != NULL, lock == olock) &&
114 ergo(lock == NULL, olock == NULL)) &&
115 ergo(ols->ols_state == OLS_CANCELLED,
116 olock == NULL && !handle_used) &&
118 * DLM lock is destroyed only after we have seen cancellation
121 ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
122 !olock->l_destroyed) &&
123 ergo(ols->ols_state == OLS_GRANTED,
125 olock->l_req_mode == olock->l_granted_mode &&
129 /*****************************************************************************
136 * Breaks a link between osc_lock and dlm_lock.
138 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
140 struct ldlm_lock *dlmlock;
142 cfs_spin_lock(&osc_ast_guard);
143 dlmlock = olck->ols_lock;
144 if (dlmlock == NULL) {
145 cfs_spin_unlock(&osc_ast_guard);
149 olck->ols_lock = NULL;
150 /* wb(); --- for all who checks (ols->ols_lock != NULL) before
151 * call to osc_lock_detach() */
152 dlmlock->l_ast_data = NULL;
153 olck->ols_handle.cookie = 0ULL;
154 cfs_spin_unlock(&osc_ast_guard);
156 lock_res_and_lock(dlmlock);
157 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
158 struct cl_object *obj = olck->ols_cl.cls_obj;
159 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
162 cl_object_attr_lock(obj);
163 /* Must get the value under the lock to avoid possible races. */
164 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
165 /* Update the kms. Need to loop all granted locks.
166 * Not a problem for the client */
167 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
169 cl_object_attr_set(env, obj, attr, CAT_KMS);
170 cl_object_attr_unlock(obj);
172 unlock_res_and_lock(dlmlock);
174 /* release a reference taken in osc_lock_upcall0(). */
175 LASSERT(olck->ols_has_ref);
176 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
177 LDLM_LOCK_RELEASE(dlmlock);
178 olck->ols_has_ref = 0;
181 static int osc_lock_unhold(struct osc_lock *ols)
187 result = osc_cancel_base(&ols->ols_handle,
188 ols->ols_einfo.ei_mode);
193 static int osc_lock_unuse(const struct lu_env *env,
194 const struct cl_lock_slice *slice)
196 struct osc_lock *ols = cl2osc_lock(slice);
198 LINVRNT(osc_lock_invariant(ols));
200 switch (ols->ols_state) {
202 LASSERT(!ols->ols_hold);
203 LASSERT(ols->ols_agl);
205 case OLS_UPCALL_RECEIVED:
206 LASSERT(!ols->ols_hold);
207 ols->ols_state = OLS_NEW;
210 LASSERT(!ols->ols_glimpse);
211 LASSERT(ols->ols_hold);
213 * Move lock into OLS_RELEASED state before calling
214 * osc_cancel_base() so that possible synchronous cancellation
215 * (that always happens e.g., for liblustre) sees that lock is
218 ols->ols_state = OLS_RELEASED;
219 return osc_lock_unhold(ols);
221 CERROR("Impossible state: %d\n", ols->ols_state);
226 static void osc_lock_fini(const struct lu_env *env,
227 struct cl_lock_slice *slice)
229 struct osc_lock *ols = cl2osc_lock(slice);
231 LINVRNT(osc_lock_invariant(ols));
233 * ->ols_hold can still be true at this point if, for example, a
234 * thread that requested a lock was killed (and released a reference
235 * to the lock), before reply from a server was received. In this case
236 * lock is destroyed immediately after upcall.
238 osc_lock_unhold(ols);
239 LASSERT(ols->ols_lock == NULL);
240 LASSERT(cfs_atomic_read(&ols->ols_pageref) == 0 ||
241 cfs_atomic_read(&ols->ols_pageref) == _PAGEREF_MAGIC);
243 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
246 void osc_lock_build_res(const struct lu_env *env, const struct osc_object *obj,
247 struct ldlm_res_id *resname)
249 const struct lu_fid *fid = lu_object_fid(&obj->oo_cl.co_lu);
252 * In the perfect world of the future, where ost servers talk
255 fid_build_reg_res_name(fid, resname);
258 * In reality, where ost server expects ->lsm_object_id and
259 * ->lsm_object_seq in rename.
261 osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_seq,
266 static void osc_lock_build_policy(const struct lu_env *env,
267 const struct cl_lock *lock,
268 ldlm_policy_data_t *policy)
270 const struct cl_lock_descr *d = &lock->cll_descr;
272 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
273 policy->l_extent.gid = d->cld_gid;
276 static int osc_enq2ldlm_flags(__u32 enqflags)
280 LASSERT((enqflags & ~CEF_MASK) == 0);
282 if (enqflags & CEF_NONBLOCK)
283 result |= LDLM_FL_BLOCK_NOWAIT;
284 if (enqflags & CEF_ASYNC)
285 result |= LDLM_FL_HAS_INTENT;
286 if (enqflags & CEF_DISCARD_DATA)
287 result |= LDLM_AST_DISCARD_DATA;
292 * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
293 * pointers. Initialized in osc_init().
295 cfs_spinlock_t osc_ast_guard;
297 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
299 struct osc_lock *olck;
301 lock_res_and_lock(dlm_lock);
302 cfs_spin_lock(&osc_ast_guard);
303 olck = dlm_lock->l_ast_data;
305 struct cl_lock *lock = olck->ols_cl.cls_lock;
307 * If osc_lock holds a reference on ldlm lock, return it even
308 * when cl_lock is in CLS_FREEING state. This way
310 * osc_ast_data_get(dlmlock) == NULL
312 * guarantees that all osc references on dlmlock were
313 * released. osc_dlm_blocking_ast0() relies on that.
315 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
316 cl_lock_get_trust(lock);
317 lu_ref_add_atomic(&lock->cll_reference,
318 "ast", cfs_current());
322 cfs_spin_unlock(&osc_ast_guard);
323 unlock_res_and_lock(dlm_lock);
327 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
329 struct cl_lock *lock;
331 lock = olck->ols_cl.cls_lock;
332 lu_ref_del(&lock->cll_reference, "ast", cfs_current());
333 cl_lock_put(env, lock);
337 * Updates object attributes from a lock value block (lvb) received together
338 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
341 * This can be optimized to not update attributes when lock is a result of a
344 * Called under lock and resource spin-locks.
346 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
350 struct cl_object *obj;
351 struct lov_oinfo *oinfo;
352 struct cl_attr *attr;
357 if (!(olck->ols_flags & LDLM_FL_LVB_READY))
360 lvb = &olck->ols_lvb;
361 obj = olck->ols_cl.cls_obj;
362 oinfo = cl2osc(obj)->oo_oinfo;
363 attr = &osc_env_info(env)->oti_attr;
364 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
365 cl_lvb2attr(attr, lvb);
367 cl_object_attr_lock(obj);
369 struct ldlm_lock *dlmlock;
372 dlmlock = olck->ols_lock;
373 LASSERT(dlmlock != NULL);
375 /* re-grab LVB from a dlm lock under DLM spin-locks. */
376 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
377 size = lvb->lvb_size;
378 /* Extend KMS up to the end of this lock and no further
379 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
380 if (size > dlmlock->l_policy_data.l_extent.end)
381 size = dlmlock->l_policy_data.l_extent.end + 1;
382 if (size >= oinfo->loi_kms) {
383 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
384 ", kms="LPU64, lvb->lvb_size, size);
386 attr->cat_kms = size;
388 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
389 LPU64"; leaving kms="LPU64", end="LPU64,
390 lvb->lvb_size, oinfo->loi_kms,
391 dlmlock->l_policy_data.l_extent.end);
393 ldlm_lock_allow_match_locked(dlmlock);
394 } else if (rc == -ENAVAIL && olck->ols_glimpse) {
395 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
396 " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
401 cl_object_attr_set(env, obj, attr, valid);
403 cl_object_attr_unlock(obj);
409 * Called when a lock is granted, from an upcall (when server returned a
410 * granted lock), or from completion AST, when server returned a blocked lock.
412 * Called under lock and resource spin-locks, that are released temporarily
415 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
416 struct ldlm_lock *dlmlock, int rc)
418 struct ldlm_extent *ext;
419 struct cl_lock *lock;
420 struct cl_lock_descr *descr;
422 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
425 if (olck->ols_state < OLS_GRANTED) {
426 lock = olck->ols_cl.cls_lock;
427 ext = &dlmlock->l_policy_data.l_extent;
428 descr = &osc_env_info(env)->oti_descr;
429 descr->cld_obj = lock->cll_descr.cld_obj;
431 /* XXX check that ->l_granted_mode is valid. */
432 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
433 descr->cld_start = cl_index(descr->cld_obj, ext->start);
434 descr->cld_end = cl_index(descr->cld_obj, ext->end);
435 descr->cld_gid = ext->gid;
437 * tell upper layers the extent of the lock that was actually
440 olck->ols_state = OLS_GRANTED;
441 osc_lock_lvb_update(env, olck, rc);
443 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
444 * to take a semaphore on a parent lock. This is safe, because
445 * spin-locks are needed to protect consistency of
446 * dlmlock->l_*_mode and LVB, and we have finished processing
448 unlock_res_and_lock(dlmlock);
449 cl_lock_modify(env, lock, descr);
450 cl_lock_signal(env, lock);
451 LINVRNT(osc_lock_invariant(olck));
452 lock_res_and_lock(dlmlock);
457 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
460 struct ldlm_lock *dlmlock;
464 dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
465 LASSERT(dlmlock != NULL);
467 lock_res_and_lock(dlmlock);
468 cfs_spin_lock(&osc_ast_guard);
469 LASSERT(dlmlock->l_ast_data == olck);
470 LASSERT(olck->ols_lock == NULL);
471 olck->ols_lock = dlmlock;
472 cfs_spin_unlock(&osc_ast_guard);
475 * Lock might be not yet granted. In this case, completion ast
476 * (osc_ldlm_completion_ast()) comes later and finishes lock
479 if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
480 osc_lock_granted(env, olck, dlmlock, 0);
481 unlock_res_and_lock(dlmlock);
484 * osc_enqueue_interpret() decrefs asynchronous locks, counter
487 ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
490 /* lock reference taken by ldlm_handle2lock_long() is owned by
491 * osc_lock and released in osc_lock_detach() */
492 lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
493 olck->ols_has_ref = 1;
497 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
498 * received from a server, or after osc_enqueue_base() matched a local DLM
501 static int osc_lock_upcall(void *cookie, int errcode)
503 struct osc_lock *olck = cookie;
504 struct cl_lock_slice *slice = &olck->ols_cl;
505 struct cl_lock *lock = slice->cls_lock;
507 struct cl_env_nest nest;
510 env = cl_env_nested_get(&nest);
514 cl_lock_mutex_get(env, lock);
516 LASSERT(lock->cll_state >= CLS_QUEUING);
517 if (olck->ols_state == OLS_ENQUEUED) {
518 olck->ols_state = OLS_UPCALL_RECEIVED;
519 rc = ldlm_error2errno(errcode);
520 } else if (olck->ols_state == OLS_CANCELLED) {
523 CERROR("Impossible state: %d\n", olck->ols_state);
527 struct ldlm_lock *dlmlock;
529 dlmlock = ldlm_handle2lock(&olck->ols_handle);
530 if (dlmlock != NULL) {
531 lock_res_and_lock(dlmlock);
532 cfs_spin_lock(&osc_ast_guard);
533 LASSERT(olck->ols_lock == NULL);
534 dlmlock->l_ast_data = NULL;
535 olck->ols_handle.cookie = 0ULL;
536 cfs_spin_unlock(&osc_ast_guard);
537 ldlm_lock_fail_match_locked(dlmlock);
538 unlock_res_and_lock(dlmlock);
539 LDLM_LOCK_PUT(dlmlock);
542 if (olck->ols_glimpse) {
543 olck->ols_glimpse = 0;
546 osc_lock_upcall0(env, olck);
549 /* Error handling, some errors are tolerable. */
550 if (olck->ols_locklessable && rc == -EUSERS) {
551 /* This is a tolerable error, turn this lock into
554 osc_object_set_contended(cl2osc(slice->cls_obj));
555 LASSERT(slice->cls_ops == &osc_lock_ops);
557 /* Change this lock to ldlmlock-less lock. */
558 osc_lock_to_lockless(env, olck, 1);
559 olck->ols_state = OLS_GRANTED;
561 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
562 osc_lock_lvb_update(env, olck, rc);
563 cl_lock_delete(env, lock);
564 /* Hide the error. */
569 cl_lock_signal(env, lock);
570 /* del user for lock upcall cookie */
571 cl_unuse_try(env, lock);
573 /* del user for lock upcall cookie */
574 cl_lock_user_del(env, lock);
575 cl_lock_error(env, lock, rc);
578 cl_lock_mutex_put(env, lock);
580 /* release cookie reference, acquired by osc_lock_enqueue() */
581 lu_ref_del(&lock->cll_reference, "upcall", lock);
582 cl_lock_put(env, lock);
584 cl_env_nested_put(&nest, env);
586 /* should never happen, similar to osc_ldlm_blocking_ast(). */
592 * Core of osc_dlm_blocking_ast() logic.
594 static void osc_lock_blocking(const struct lu_env *env,
595 struct ldlm_lock *dlmlock,
596 struct osc_lock *olck, int blocking)
598 struct cl_lock *lock = olck->ols_cl.cls_lock;
600 LASSERT(olck->ols_lock == dlmlock);
601 CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
602 LASSERT(!osc_lock_is_lockless(olck));
605 * Lock might be still addref-ed here, if e.g., blocking ast
606 * is sent for a failed lock.
608 osc_lock_unhold(olck);
610 if (blocking && olck->ols_state < OLS_BLOCKED)
612 * Move osc_lock into OLS_BLOCKED before canceling the lock,
613 * because it recursively re-enters osc_lock_blocking(), with
614 * the state set to OLS_CANCELLED.
616 olck->ols_state = OLS_BLOCKED;
618 * cancel and destroy lock at least once no matter how blocking ast is
619 * entered (see comment above osc_ldlm_blocking_ast() for use
620 * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
622 cl_lock_cancel(env, lock);
623 cl_lock_delete(env, lock);
627 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
628 * and ldlm_lock caches.
630 static int osc_dlm_blocking_ast0(const struct lu_env *env,
631 struct ldlm_lock *dlmlock,
632 void *data, int flag)
634 struct osc_lock *olck;
635 struct cl_lock *lock;
639 LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
642 olck = osc_ast_data_get(dlmlock);
644 lock = olck->ols_cl.cls_lock;
645 cl_lock_mutex_get(env, lock);
646 LINVRNT(osc_lock_invariant(olck));
647 if (olck->ols_ast_wait) {
648 /* wake up osc_lock_use() */
649 cl_lock_signal(env, lock);
650 olck->ols_ast_wait = 0;
653 * Lock might have been canceled while this thread was
654 * sleeping for lock mutex, but olck is pinned in memory.
656 if (olck == dlmlock->l_ast_data) {
658 * NOTE: DLM sends blocking AST's for failed locks
659 * (that are still in pre-OLS_GRANTED state)
660 * too, and they have to be canceled otherwise
661 * DLM lock is never destroyed and stuck in
664 * Alternatively, ldlm_cli_cancel() can be
665 * called here directly for osc_locks with
666 * ols_state < OLS_GRANTED to maintain an
667 * invariant that ->clo_cancel() is only called
668 * for locks that were granted.
670 LASSERT(data == olck);
671 osc_lock_blocking(env, dlmlock,
672 olck, flag == LDLM_CB_BLOCKING);
675 cl_lock_mutex_put(env, lock);
676 osc_ast_data_put(env, olck);
679 * DLM lock exists, but there is no cl_lock attached to it.
680 * This is a `normal' race. cl_object and its cl_lock's can be
681 * removed by memory pressure, together with all pages.
683 cancel = (flag == LDLM_CB_BLOCKING);
686 struct lustre_handle *lockh;
688 lockh = &osc_env_info(env)->oti_handle;
689 ldlm_lock2handle(dlmlock, lockh);
690 result = ldlm_cli_cancel(lockh);
697 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
698 * some other lock, or is canceled. This function is installed as a
699 * ldlm_lock::l_blocking_ast() for client extent locks.
701 * Control flow is tricky, because ldlm uses the same call-back
702 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
704 * \param dlmlock lock for which ast occurred.
706 * \param new description of a conflicting lock in case of blocking ast.
708 * \param data value of dlmlock->l_ast_data
710 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
711 * cancellation and blocking ast's.
713 * Possible use cases:
715 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
716 * lock due to lock lru pressure, or explicit user request to purge
719 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
720 * us that dlmlock conflicts with another lock that some client is
721 * enqueing. Lock is canceled.
723 * - cl_lock_cancel() is called. osc_lock_cancel() calls
724 * ldlm_cli_cancel() that calls
726 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
728 * recursively entering osc_ldlm_blocking_ast().
730 * - client cancels lock voluntary (e.g., as a part of early cancellation):
733 * osc_lock_cancel()->
734 * ldlm_cli_cancel()->
735 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
738 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
739 struct ldlm_lock_desc *new, void *data,
743 struct cl_env_nest nest;
747 * This can be called in the context of outer IO, e.g.,
750 * ->osc_enqueue_base()->...
751 * ->ldlm_prep_elc_req()->...
752 * ->ldlm_cancel_callback()->...
753 * ->osc_ldlm_blocking_ast()
755 * new environment has to be created to not corrupt outer context.
757 env = cl_env_nested_get(&nest);
759 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
760 cl_env_nested_put(&nest, env);
762 result = PTR_ERR(env);
764 * XXX This should never happen, as cl_lock is
765 * stuck. Pre-allocated environment a la vvp_inode_fini_env
771 if (result == -ENODATA)
774 CERROR("BAST failed: %d\n", result);
779 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
780 int flags, void *data)
782 struct cl_env_nest nest;
784 struct osc_lock *olck;
785 struct cl_lock *lock;
789 /* first, do dlm part of the work */
790 dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
791 /* then, notify cl_lock */
792 env = cl_env_nested_get(&nest);
794 olck = osc_ast_data_get(dlmlock);
796 lock = olck->ols_cl.cls_lock;
797 cl_lock_mutex_get(env, lock);
799 * ldlm_handle_cp_callback() copied LVB from request
800 * to lock->l_lvb_data, store it in osc_lock.
802 LASSERT(dlmlock->l_lvb_data != NULL);
803 lock_res_and_lock(dlmlock);
804 olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
805 if (olck->ols_lock == NULL) {
807 * upcall (osc_lock_upcall()) hasn't yet been
808 * called. Do nothing now, upcall will bind
809 * olck to dlmlock and signal the waiters.
811 * This maintains an invariant that osc_lock
812 * and ldlm_lock are always bound when
813 * osc_lock is in OLS_GRANTED state.
815 } else if (dlmlock->l_granted_mode ==
816 dlmlock->l_req_mode) {
817 osc_lock_granted(env, olck, dlmlock, dlmrc);
819 unlock_res_and_lock(dlmlock);
822 CL_LOCK_DEBUG(D_ERROR, env, lock,
823 "dlmlock returned %d\n", dlmrc);
824 cl_lock_error(env, lock, dlmrc);
826 cl_lock_mutex_put(env, lock);
827 osc_ast_data_put(env, olck);
830 result = -ELDLM_NO_LOCK_DATA;
831 cl_env_nested_put(&nest, env);
833 result = PTR_ERR(env);
834 return dlmrc ?: result;
837 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
839 struct ptlrpc_request *req = data;
840 struct osc_lock *olck;
841 struct cl_lock *lock;
842 struct cl_object *obj;
843 struct cl_env_nest nest;
846 struct req_capsule *cap;
849 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
851 env = cl_env_nested_get(&nest);
854 * osc_ast_data_get() has to go after environment is
855 * allocated, because osc_ast_data() acquires a
856 * reference to a lock, and it can only be released in
859 olck = osc_ast_data_get(dlmlock);
861 lock = olck->ols_cl.cls_lock;
862 cl_lock_mutex_get(env, lock);
864 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
865 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
867 result = req_capsule_server_pack(cap);
869 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
870 obj = lock->cll_descr.cld_obj;
871 result = cl_object_glimpse(env, obj, lvb);
873 cl_lock_mutex_put(env, lock);
874 osc_ast_data_put(env, olck);
877 * These errors are normal races, so we don't want to
878 * fill the console with messages by calling
881 lustre_pack_reply(req, 1, NULL, NULL);
882 result = -ELDLM_NO_LOCK_DATA;
884 cl_env_nested_put(&nest, env);
886 result = PTR_ERR(env);
887 req->rq_status = result;
891 static unsigned long osc_lock_weigh(const struct lu_env *env,
892 const struct cl_lock_slice *slice)
895 * don't need to grab coh_page_guard since we don't care the exact #
898 return cl_object_header(slice->cls_obj)->coh_pages;
902 * Get the weight of dlm lock for early cancellation.
904 * XXX: it should return the pages covered by this \a dlmlock.
906 static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
908 struct cl_env_nest nest;
910 struct osc_lock *lock;
912 unsigned long weight;
917 * osc_ldlm_weigh_ast has a complex context since it might be called
918 * because of lock canceling, or from user's input. We have to make
919 * a new environment for it. Probably it is implementation safe to use
920 * the upper context because cl_lock_put don't modify environment
921 * variables. But in case of ..
923 env = cl_env_nested_get(&nest);
925 /* Mostly because lack of memory, tend to eliminate this lock*/
928 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
929 lock = osc_ast_data_get(dlmlock);
931 /* cl_lock was destroyed because of memory pressure.
932 * It is much reasonable to assign this type of lock
935 GOTO(out, weight = 0);
938 cll = lock->ols_cl.cls_lock;
939 cl_lock_mutex_get(env, cll);
940 weight = cl_lock_weigh(env, cll);
941 cl_lock_mutex_put(env, cll);
942 osc_ast_data_put(env, lock);
946 cl_env_nested_put(&nest, env);
950 static void osc_lock_build_einfo(const struct lu_env *env,
951 const struct cl_lock *clock,
952 struct osc_lock *lock,
953 struct ldlm_enqueue_info *einfo)
955 enum cl_lock_mode mode;
957 mode = clock->cll_descr.cld_mode;
958 if (mode == CLM_PHANTOM)
960 * For now, enqueue all glimpse locks in read mode. In the
961 * future, client might choose to enqueue LCK_PW lock for
962 * glimpse on a file opened for write.
966 einfo->ei_type = LDLM_EXTENT;
967 einfo->ei_mode = osc_cl_lock2ldlm(mode);
968 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
969 einfo->ei_cb_cp = osc_ldlm_completion_ast;
970 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
971 einfo->ei_cb_wg = osc_ldlm_weigh_ast;
972 einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
976 * Determine if the lock should be converted into a lockless lock.
979 * - if the lock has an explicite requirment for a non-lockless lock;
980 * - if the io lock request type ci_lockreq;
981 * - send the enqueue rpc to ost to make the further decision;
982 * - special treat to truncate lockless lock
984 * Additional policy can be implemented here, e.g., never do lockless-io
987 static void osc_lock_to_lockless(const struct lu_env *env,
988 struct osc_lock *ols, int force)
990 struct cl_lock_slice *slice = &ols->ols_cl;
991 struct cl_lock *lock = slice->cls_lock;
993 LASSERT(ols->ols_state == OLS_NEW ||
994 ols->ols_state == OLS_UPCALL_RECEIVED);
997 ols->ols_locklessable = 1;
998 LASSERT(cl_lock_is_mutexed(lock));
999 slice->cls_ops = &osc_lock_lockless_ops;
1001 struct osc_io *oio = osc_env_io(env);
1002 struct cl_io *io = oio->oi_cl.cis_io;
1003 struct cl_object *obj = slice->cls_obj;
1004 struct osc_object *oob = cl2osc(obj);
1005 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
1006 struct obd_connect_data *ocd;
1008 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
1009 io->ci_lockreq == CILR_MAYBE ||
1010 io->ci_lockreq == CILR_NEVER);
1012 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
1013 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
1014 (io->ci_lockreq == CILR_MAYBE) &&
1015 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
1016 if (io->ci_lockreq == CILR_NEVER ||
1018 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
1019 /* lockless truncate */
1020 (cl_io_is_trunc(io) &&
1021 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1022 osd->od_lockless_truncate)) {
1023 ols->ols_locklessable = 1;
1024 slice->cls_ops = &osc_lock_lockless_ops;
1027 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1030 static int osc_lock_compatible(const struct osc_lock *qing,
1031 const struct osc_lock *qed)
1033 enum cl_lock_mode qing_mode;
1034 enum cl_lock_mode qed_mode;
1036 qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode;
1037 if (qed->ols_glimpse &&
1038 (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ))
1041 qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode;
1042 return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ));
1046 * Cancel all conflicting locks and wait for them to be destroyed.
1048 * This function is used for two purposes:
1050 * - early cancel all conflicting locks before starting IO, and
1052 * - guarantee that pages added to the page cache by lockless IO are never
1053 * covered by locks other than lockless IO lock, and, hence, are not
1054 * visible to other threads.
1056 static int osc_lock_enqueue_wait(const struct lu_env *env,
1057 const struct osc_lock *olck)
1059 struct cl_lock *lock = olck->ols_cl.cls_lock;
1060 struct cl_lock_descr *descr = &lock->cll_descr;
1061 struct cl_object_header *hdr = cl_object_header(descr->cld_obj);
1062 struct cl_lock *scan;
1063 struct cl_lock *conflict= NULL;
1064 int lockless = osc_lock_is_lockless(olck);
1068 LASSERT(cl_lock_is_mutexed(lock));
1070 /* make it enqueue anyway for glimpse lock, because we actually
1071 * don't need to cancel any conflicting locks. */
1072 if (olck->ols_glimpse)
1075 cfs_spin_lock(&hdr->coh_lock_guard);
1076 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
1077 struct cl_lock_descr *cld = &scan->cll_descr;
1078 const struct osc_lock *scan_ols;
1083 if (scan->cll_state < CLS_QUEUING ||
1084 scan->cll_state == CLS_FREEING ||
1085 cld->cld_start > descr->cld_end ||
1086 cld->cld_end < descr->cld_start)
1089 /* overlapped and living locks. */
1091 /* We're not supposed to give up group lock. */
1092 if (scan->cll_descr.cld_mode == CLM_GROUP) {
1093 LASSERT(descr->cld_mode != CLM_GROUP ||
1094 descr->cld_gid != scan->cll_descr.cld_gid);
1098 scan_ols = osc_lock_at(scan);
1100 /* We need to cancel the compatible locks if we're enqueuing
1101 * a lockless lock, for example:
1102 * imagine that client has PR lock on [0, 1000], and thread T0
1103 * is doing lockless IO in [500, 1500] region. Concurrent
1104 * thread T1 can see lockless data in [500, 1000], which is
1105 * wrong, because these data are possibly stale. */
1106 if (!lockless && osc_lock_compatible(olck, scan_ols))
1109 /* Now @scan is conflicting with @lock, this means current
1110 * thread have to sleep for @scan being destroyed. */
1111 if (scan_ols->ols_owner == osc_env_io(env)) {
1112 CERROR("DEADLOCK POSSIBLE!\n");
1113 CL_LOCK_DEBUG(D_ERROR, env, scan, "queued.\n");
1114 CL_LOCK_DEBUG(D_ERROR, env, lock, "queuing.\n");
1115 libcfs_debug_dumpstack(NULL);
1117 cl_lock_get_trust(scan);
1121 cfs_spin_unlock(&hdr->coh_lock_guard);
1124 if (lock->cll_descr.cld_mode == CLM_GROUP) {
1125 /* we want a group lock but a previous lock request
1126 * conflicts, we do not wait but return 0 so the
1127 * request is send to the server
1129 CDEBUG(D_DLMTRACE, "group lock %p is conflicted "
1130 "with %p, no wait, send to server\n",
1132 cl_lock_put(env, conflict);
1135 CDEBUG(D_DLMTRACE, "lock %p is conflicted with %p, "
1138 LASSERT(lock->cll_conflict == NULL);
1139 lu_ref_add(&conflict->cll_reference, "cancel-wait",
1141 lock->cll_conflict = conflict;
1149 * Implementation of cl_lock_operations::clo_enqueue() method for osc
1150 * layer. This initiates ldlm enqueue:
1152 * - cancels conflicting locks early (osc_lock_enqueue_wait());
1154 * - calls osc_enqueue_base() to do actual enqueue.
1156 * osc_enqueue_base() is supplied with an upcall function that is executed
1157 * when lock is received either after a local cached ldlm lock is matched, or
1158 * when a reply from the server is received.
1160 * This function does not wait for the network communication to complete.
1162 static int osc_lock_enqueue(const struct lu_env *env,
1163 const struct cl_lock_slice *slice,
1164 struct cl_io *unused, __u32 enqflags)
1166 struct osc_lock *ols = cl2osc_lock(slice);
1167 struct cl_lock *lock = ols->ols_cl.cls_lock;
1171 LASSERT(cl_lock_is_mutexed(lock));
1172 LASSERTF(ols->ols_state == OLS_NEW,
1173 "Impossible state: %d\n", ols->ols_state);
1175 ols->ols_flags = osc_enq2ldlm_flags(enqflags);
1176 if (enqflags & CEF_AGL) {
1177 ols->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1180 if (ols->ols_flags & LDLM_FL_HAS_INTENT)
1181 ols->ols_glimpse = 1;
1182 if (!osc_lock_is_lockless(ols) && !(enqflags & CEF_MUST))
1183 /* try to convert this lock to a lockless lock */
1184 osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER));
1186 result = osc_lock_enqueue_wait(env, ols);
1188 if (!osc_lock_is_lockless(ols)) {
1189 struct osc_object *obj = cl2osc(slice->cls_obj);
1190 struct osc_thread_info *info = osc_env_info(env);
1191 struct ldlm_res_id *resname = &info->oti_resname;
1192 ldlm_policy_data_t *policy = &info->oti_policy;
1193 struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1195 if (ols->ols_locklessable)
1196 ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1198 /* a reference for lock, passed as an upcall cookie */
1200 lu_ref_add(&lock->cll_reference, "upcall", lock);
1201 /* a user for lock also */
1202 cl_lock_user_add(env, lock);
1203 ols->ols_state = OLS_ENQUEUED;
1206 * XXX: this is possible blocking point as
1207 * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1210 osc_lock_build_res(env, obj, resname);
1211 osc_lock_build_policy(env, lock, policy);
1212 result = osc_enqueue_base(osc_export(obj), resname,
1213 &ols->ols_flags, policy,
1215 obj->oo_oinfo->loi_kms_valid,
1217 ols, einfo, &ols->ols_handle,
1218 PTLRPCD_SET, 1, ols->ols_agl);
1220 cl_lock_user_del(env, lock);
1221 lu_ref_del(&lock->cll_reference,
1223 cl_lock_put(env, lock);
1224 if (unlikely(result == -ECANCELED)) {
1225 ols->ols_state = OLS_NEW;
1230 ols->ols_state = OLS_GRANTED;
1231 ols->ols_owner = osc_env_io(env);
1234 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1238 static int osc_lock_wait(const struct lu_env *env,
1239 const struct cl_lock_slice *slice)
1241 struct osc_lock *olck = cl2osc_lock(slice);
1242 struct cl_lock *lock = olck->ols_cl.cls_lock;
1244 LINVRNT(osc_lock_invariant(olck));
1246 if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) {
1247 if (olck->ols_flags & LDLM_FL_LVB_READY) {
1249 } else if (olck->ols_agl) {
1250 olck->ols_state = OLS_NEW;
1252 LASSERT(lock->cll_error);
1253 return lock->cll_error;
1257 if (olck->ols_state == OLS_NEW) {
1258 if (lock->cll_descr.cld_enq_flags & CEF_NO_REENQUEUE) {
1263 LASSERT(olck->ols_agl);
1265 rc = osc_lock_enqueue(env, slice, NULL, CEF_ASYNC |
1270 return CLO_REENQUEUED;
1274 LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1275 lock->cll_error == 0, olck->ols_lock != NULL));
1277 return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1281 * An implementation of cl_lock_operations::clo_use() method that pins cached
1284 static int osc_lock_use(const struct lu_env *env,
1285 const struct cl_lock_slice *slice)
1287 struct osc_lock *olck = cl2osc_lock(slice);
1290 LASSERT(!olck->ols_hold);
1293 * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1294 * flag is not set. This protects us from a concurrent blocking ast.
1296 rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1299 olck->ols_state = OLS_GRANTED;
1301 struct cl_lock *lock;
1304 * Lock is being cancelled somewhere within
1305 * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1306 * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1309 lock = slice->cls_lock;
1310 LASSERT(lock->cll_state == CLS_INTRANSIT);
1311 LASSERT(lock->cll_users > 0);
1312 /* set a flag for osc_dlm_blocking_ast0() to signal the
1314 olck->ols_ast_wait = 1;
1320 static int osc_lock_flush(struct osc_lock *ols, int discard)
1322 struct cl_lock *lock = ols->ols_cl.cls_lock;
1323 struct cl_env_nest nest;
1327 env = cl_env_nested_get(&nest);
1329 result = cl_lock_page_out(env, lock, discard);
1330 cl_env_nested_put(&nest, env);
1332 result = PTR_ERR(env);
1335 LINVRNT(!osc_lock_has_pages(ols));
1341 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1342 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1343 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1344 * with some other lock some where in the cluster. This function does the
1347 * - invalidates all pages protected by this lock (after sending dirty
1348 * ones to the server, as necessary);
1350 * - decref's underlying ldlm lock;
1352 * - cancels ldlm lock (ldlm_cli_cancel()).
1354 static void osc_lock_cancel(const struct lu_env *env,
1355 const struct cl_lock_slice *slice)
1357 struct cl_lock *lock = slice->cls_lock;
1358 struct osc_lock *olck = cl2osc_lock(slice);
1359 struct ldlm_lock *dlmlock = olck->ols_lock;
1363 LASSERT(cl_lock_is_mutexed(lock));
1364 LINVRNT(osc_lock_invariant(olck));
1366 if (dlmlock != NULL) {
1369 discard = !!(dlmlock->l_flags & LDLM_FL_DISCARD_DATA);
1370 result = osc_lock_flush(olck, discard);
1371 osc_lock_unhold(olck);
1373 lock_res_and_lock(dlmlock);
1374 /* Now that we're the only user of dlm read/write reference,
1375 * mostly the ->l_readers + ->l_writers should be zero.
1376 * However, there is a corner case.
1377 * See bug 18829 for details.*/
1378 do_cancel = (dlmlock->l_readers == 0 &&
1379 dlmlock->l_writers == 0);
1380 dlmlock->l_flags |= LDLM_FL_CBPENDING;
1381 unlock_res_and_lock(dlmlock);
1383 result = ldlm_cli_cancel(&olck->ols_handle);
1385 CL_LOCK_DEBUG(D_ERROR, env, lock,
1386 "lock %p cancel failure with error(%d)\n",
1389 olck->ols_state = OLS_CANCELLED;
1390 olck->ols_flags &= ~LDLM_FL_LVB_READY;
1391 osc_lock_detach(env, olck);
1394 #ifdef INVARIANT_CHECK
1395 static int check_cb(const struct lu_env *env, struct cl_io *io,
1396 struct cl_page *page, void *cbdata)
1398 struct cl_lock *lock = cbdata;
1400 if (lock->cll_descr.cld_mode == CLM_READ) {
1401 struct cl_lock *tmp;
1402 tmp = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1405 cl_lock_put(env, tmp);
1406 return CLP_GANG_OKAY;
1410 CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1411 CL_PAGE_DEBUG(D_ERROR, env, page, "\n");
1412 return CLP_GANG_ABORT;
1416 * Returns true iff there are pages under \a olck not protected by other
1419 static int osc_lock_has_pages(struct osc_lock *olck)
1421 struct cl_lock *lock;
1422 struct cl_lock_descr *descr;
1423 struct cl_object *obj;
1424 struct osc_object *oob;
1425 struct cl_env_nest nest;
1430 env = cl_env_nested_get(&nest);
1434 obj = olck->ols_cl.cls_obj;
1436 io = &oob->oo_debug_io;
1437 lock = olck->ols_cl.cls_lock;
1438 descr = &lock->cll_descr;
1440 cfs_mutex_lock(&oob->oo_debug_mutex);
1442 io->ci_obj = cl_object_top(obj);
1443 cl_io_init(env, io, CIT_MISC, io->ci_obj);
1445 result = cl_page_gang_lookup(env, obj, io,
1446 descr->cld_start, descr->cld_end,
1447 check_cb, (void *)lock);
1448 if (result == CLP_GANG_ABORT)
1450 if (result == CLP_GANG_RESCHED)
1452 } while (result != CLP_GANG_OKAY);
1453 cl_io_fini(env, io);
1454 cfs_mutex_unlock(&oob->oo_debug_mutex);
1455 cl_env_nested_put(&nest, env);
1457 return (result == CLP_GANG_ABORT);
1460 static int osc_lock_has_pages(struct osc_lock *olck)
1464 #endif /* INVARIANT_CHECK */
1466 static void osc_lock_delete(const struct lu_env *env,
1467 const struct cl_lock_slice *slice)
1469 struct osc_lock *olck;
1471 olck = cl2osc_lock(slice);
1472 if (olck->ols_glimpse) {
1473 LASSERT(!olck->ols_hold);
1474 LASSERT(!olck->ols_lock);
1478 LINVRNT(osc_lock_invariant(olck));
1479 LINVRNT(!osc_lock_has_pages(olck));
1481 osc_lock_unhold(olck);
1482 osc_lock_detach(env, olck);
1486 * Implements cl_lock_operations::clo_state() method for osc layer.
1488 * Maintains osc_lock::ols_owner field.
1490 * This assumes that lock always enters CLS_HELD (from some other state) in
1491 * the same IO context as one that requested the lock. This should not be a
1492 * problem, because context is by definition shared by all activity pertaining
1493 * to the same high-level IO.
1495 static void osc_lock_state(const struct lu_env *env,
1496 const struct cl_lock_slice *slice,
1497 enum cl_lock_state state)
1499 struct osc_lock *lock = cl2osc_lock(slice);
1502 * XXX multiple io contexts can use the lock at the same time.
1504 LINVRNT(osc_lock_invariant(lock));
1505 if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1506 struct osc_io *oio = osc_env_io(env);
1508 LASSERT(lock->ols_owner == NULL);
1509 lock->ols_owner = oio;
1510 } else if (state != CLS_HELD)
1511 lock->ols_owner = NULL;
1514 static int osc_lock_print(const struct lu_env *env, void *cookie,
1515 lu_printer_t p, const struct cl_lock_slice *slice)
1517 struct osc_lock *lock = cl2osc_lock(slice);
1520 * XXX print ldlm lock and einfo properly.
1522 (*p)(env, cookie, "%p %08x "LPX64" %d %p ",
1523 lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1524 lock->ols_state, lock->ols_owner);
1525 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1529 static int osc_lock_fits_into(const struct lu_env *env,
1530 const struct cl_lock_slice *slice,
1531 const struct cl_lock_descr *need,
1532 const struct cl_io *io)
1534 struct osc_lock *ols = cl2osc_lock(slice);
1536 if (need->cld_enq_flags & CEF_NEVER)
1539 if (need->cld_mode == CLM_PHANTOM) {
1541 return !(ols->ols_state > OLS_RELEASED);
1544 * Note: the QUEUED lock can't be matched here, otherwise
1545 * it might cause the deadlocks.
1547 * P1: enqueued read lock, create sublock1
1548 * P2: enqueued write lock, create sublock2(conflicted
1550 * P1: Grant read lock.
1551 * P1: enqueued glimpse lock(with holding sublock1_read),
1552 * matched with sublock2, waiting sublock2 to be granted.
1553 * But sublock2 can not be granted, because P1
1554 * will not release sublock1. Bang!
1556 if (ols->ols_state < OLS_GRANTED ||
1557 ols->ols_state > OLS_RELEASED)
1559 } else if (need->cld_enq_flags & CEF_MUST) {
1561 * If the lock hasn't ever enqueued, it can't be matched
1562 * because enqueue process brings in many information
1563 * which can be used to determine things such as lockless,
1566 if (ols->ols_state < OLS_UPCALL_RECEIVED &&
1567 ols->ols_locklessable)
1573 static const struct cl_lock_operations osc_lock_ops = {
1574 .clo_fini = osc_lock_fini,
1575 .clo_enqueue = osc_lock_enqueue,
1576 .clo_wait = osc_lock_wait,
1577 .clo_unuse = osc_lock_unuse,
1578 .clo_use = osc_lock_use,
1579 .clo_delete = osc_lock_delete,
1580 .clo_state = osc_lock_state,
1581 .clo_cancel = osc_lock_cancel,
1582 .clo_weigh = osc_lock_weigh,
1583 .clo_print = osc_lock_print,
1584 .clo_fits_into = osc_lock_fits_into,
1587 static int osc_lock_lockless_unuse(const struct lu_env *env,
1588 const struct cl_lock_slice *slice)
1590 struct osc_lock *ols = cl2osc_lock(slice);
1591 struct cl_lock *lock = slice->cls_lock;
1593 LASSERT(ols->ols_state == OLS_GRANTED);
1594 LINVRNT(osc_lock_invariant(ols));
1596 cl_lock_cancel(env, lock);
1597 cl_lock_delete(env, lock);
1601 static void osc_lock_lockless_cancel(const struct lu_env *env,
1602 const struct cl_lock_slice *slice)
1604 struct osc_lock *ols = cl2osc_lock(slice);
1607 result = osc_lock_flush(ols, 0);
1609 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1611 ols->ols_state = OLS_CANCELLED;
1614 static int osc_lock_lockless_wait(const struct lu_env *env,
1615 const struct cl_lock_slice *slice)
1617 struct osc_lock *olck = cl2osc_lock(slice);
1618 struct cl_lock *lock = olck->ols_cl.cls_lock;
1620 LINVRNT(osc_lock_invariant(olck));
1621 LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1623 return lock->cll_error;
1626 static void osc_lock_lockless_state(const struct lu_env *env,
1627 const struct cl_lock_slice *slice,
1628 enum cl_lock_state state)
1630 struct osc_lock *lock = cl2osc_lock(slice);
1632 LINVRNT(osc_lock_invariant(lock));
1633 if (state == CLS_HELD) {
1634 struct osc_io *oio = osc_env_io(env);
1636 LASSERT(ergo(lock->ols_owner, lock->ols_owner == oio));
1637 lock->ols_owner = oio;
1639 /* set the io to be lockless if this lock is for io's
1641 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1642 oio->oi_lockless = 1;
1646 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1647 const struct cl_lock_slice *slice,
1648 const struct cl_lock_descr *need,
1649 const struct cl_io *io)
1651 struct osc_lock *lock = cl2osc_lock(slice);
1653 if (!(need->cld_enq_flags & CEF_NEVER))
1656 /* lockless lock should only be used by its owning io. b22147 */
1657 return (lock->ols_owner == osc_env_io(env));
1660 static const struct cl_lock_operations osc_lock_lockless_ops = {
1661 .clo_fini = osc_lock_fini,
1662 .clo_enqueue = osc_lock_enqueue,
1663 .clo_wait = osc_lock_lockless_wait,
1664 .clo_unuse = osc_lock_lockless_unuse,
1665 .clo_state = osc_lock_lockless_state,
1666 .clo_fits_into = osc_lock_lockless_fits_into,
1667 .clo_cancel = osc_lock_lockless_cancel,
1668 .clo_print = osc_lock_print
1671 int osc_lock_init(const struct lu_env *env,
1672 struct cl_object *obj, struct cl_lock *lock,
1673 const struct cl_io *unused)
1675 struct osc_lock *clk;
1678 OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO);
1680 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1681 cfs_atomic_set(&clk->ols_pageref, 0);
1682 clk->ols_state = OLS_NEW;
1683 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);
1690 int osc_dlm_lock_pageref(struct ldlm_lock *dlm)
1692 struct osc_lock *olock;
1695 cfs_spin_lock(&osc_ast_guard);
1696 olock = dlm->l_ast_data;
1698 * there's a very rare race with osc_page_addref_lock(), but that
1699 * doesn't matter because in the worst case we don't cancel a lock
1700 * which we actually can, that's no harm.
1702 if (olock != NULL &&
1703 cfs_atomic_add_return(_PAGEREF_MAGIC,
1704 &olock->ols_pageref) != _PAGEREF_MAGIC) {
1705 cfs_atomic_sub(_PAGEREF_MAGIC, &olock->ols_pageref);
1708 cfs_spin_unlock(&osc_ast_guard);